Hire the Best CUDA Developers
in India

I help businesses build production-grade AI systems that are fast, scalable, and engineered for real-world use. With over 10 years of experience in software engineering, I specialize in designing AI infrastructure, CUDA-accelerated applications, Retrieval-Augmented Generation (RAG) systems, AI agents, and high-performance backend platforms that power reliable AI products. Core Areas of Expertise • CUDA & GPU Programming • AI Infrastructure Engineering • Rust Systems Programming • Python Development • Retrieval-Augmented Generation (RAG) • AI Agents & Multi-Agent Systems • LLM Applications & Orchestration • AI Inference Optimization • Vector Databases (Pinecone, ChromaDB, FAISS, Weaviate) • LangChain & LlamaIndex • Distributed Systems • Backend APIs & Microservices • Parallel Computing • Performance Optimization Every System I Build Is Designed To Be • High-performance • Scalable • Efficient • Reliable • Secure • Production-ready From document ingestion pipelines and embedding generation to semantic search, vector indexing, AI inference, GPU optimization, and cloud deployment, I build end-to-end AI systems that are designed for production rather than proof-of-concept. My focus is always on delivering solutions that remain maintainable, observable, and performant as they scale. Whether you need to build an enterprise RAG platform, optimize GPU workloads with CUDA, develop high-performance services in Rust, create intelligent AI agents, or architect scalable AI infrastructure from the ground up, I can help transform your ideas into reliable production systems. I believe great AI products are built on strong engineering foundations. That is why I focus not only on implementing AI capabilities but also on designing the infrastructure, architecture, and backend systems that ensure those solutions are fast, resilient, and ready for real-world workloads. If you are looking for an engineer who understands both modern AI and the systems that power it, send me a message, and let’s discuss how we can make it happen!

🚀Principal AI/ML Engineer — 13 years · Expert-Vetted (Top 1% on Upwork) I build production AI systems end to end — from the bare-metal GPU infrastructure all the way up to the product your users actually touch. Not prototypes that die in a notebook: real, deployed, monitored systems that hold up under load. For 13 years I've designed, trained and shipped AI across large language models, generative AI, computer vision, speech and NLP — for clients in insurance, healthcare, finance, public safety and media. Today I run the on-prem GPU infrastructure behind a national AI platform, serving a 753-billion-parameter language model (GLM-5.2) across a 32-GPU NVIDIA B200 cluster that handles 1,000+ users at once. ⭐ What I can do for you 🔹 LLM serving & cost optimization — Kill your per-token API bill. Self-host open models (GLM, Qwen, Llama, Mistral, Kimi) on your hardware or cloud with vLLM/SGLang: tensor and expert parallelism, FP8 quantization, high concurrency, low latency, predictable cost. 🔹 Fine-tuning & custom models — SFT, LoRA / QLoRA and full fine-tunes tailored to your domain and language, including hard low-resource languages. (I fine-tuned Whisper for Uzbek to 10.5% WER and built custom neural TTS voices from scratch.) 🔹 RAG & knowledge systems — RAG with hybrid dense+sparse retrieval, rerankers, vector and graph DBs (Qdrant, Pinecone, Neo4j, NebulaGraph), plus evaluation so you can prove it got better. 🔹 Computer vision — Vision and speech, not just chatbots. Face analytics, object detection and tracking, OCR and document pipelines, health-from-video (heart rate, SpO2, BMI), real-time multi-camera video, streaming voice agents, AI calling systems, voice cloning, image and video generation. 🔹 AI agents & automation — Multi-agent systems, document processing, LLM-powered classification and extraction, and end-to-end workflow automation. 🔹 Generative AI — Image and video generation, avatars, voice cloning, text-to-speech and speech-to-text. GPU infrastructure that stays up. Bare-metal and cloud clusters, CUDA, multi-node inference, autoscaling, load balancing, monitoring, CI/CD. The layer most freelancers leave you holding. 🚦 HOW ENGAGEMENTS START • Cost audit — I look at what you spend on tokens or cloud GPUs and tell you what self-hosting or a smaller fine-tune would actually cost. • Prototype-to-production rescue — it works in a notebook or a demo; I make it survive real users, real load, real uptime. • Build from zero — model choice, infra, backend, deployment, monitoring, handover docs. ✳️ HOW I WORK • We agree on the target first: latency, cost per request, accuracy, concurrency. Numbers, not vibes. • I build inside your environment: your cloud or your hardware, your repo, your keys. You own all of it. • You get infrastructure-as-code, a runbook and working monitoring, not a model file and good luck. • You talk to the person writing the code. No account manager, no team behind me. • Straight answers. If a smaller model, cheaper hardware or no AI at all is the right call, I say so before you spend. 🛠️ STACK LLM/RAG: GPT-4.1/4o/o3, Claude 4, Gemini 2.5, Llama 3.x, Qwen 3, DeepSeek, Mistral, GLM, fine-tuning, hybrid and semantic search, prompt and context engineering, structured output, knowledge-graph RAG Agents: LangGraph, LangChain, LlamaIndex, CrewAI, AutoGen, OpenAI Agents SDK, DSPy, Haystack, MCP, function calling, long-term memory, human-in-the-loop Serving & GPU: vLLM, SGLang, Ollama, NVIDIA NIM, CUDA, quantization, autoscaling, load balancing, cluster ops ML: PyTorch, TensorFlow, Hugging Face, Transformers Speech: Whisper, Faster-Whisper, speech recognition, TTS, speaker recognition, real-time voice Vision & generative AI: YOLO, ViT, OCR, segmentation, medical imaging, multimodal, image and video generation, avatars Backend: Python, FastAPI, Django, Flask, REST/GraphQL/WebSockets/gRPC, Celery, Redis, Kafka, RabbitMQ, Airflow Data: PostgreSQL, MySQL, MongoDB, Elasticsearch, Neo4j, Pinecone, Weaviate, Qdrant, Milvus, ChromaDB, FAISS Infra: AWS, GCP, Azure, RunPod, Hetzner, Vast.ai, Docker, Kubernetes, Helm, NGINX, Traefik, GitHub Actions, Linux Integrations: OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Twilio, WhatsApp Business, Slack, Salesforce, HubSpot, Stripe, OAuth2/JWT 📦 TYPICAL BUILDS Enterprise assistants and internal copilots · RAG and knowledge platforms · support, sales and SDR bots Voice assistants and AI calling · document AI and OCR pipelines · medical and financial AI Predictive analytics and recommenders · computer vision apps · workflow automation · AI dashboards 📩 Send me a message with what you're building and where it's stuck. One paragraph is enough. I'll tell you honestly whether I can help and exactly how I'd approach it, before you spend a dollar. If I'm not the right person, I'll say so.

If you can think of a problem, I can engineer the solution. My career has been built on tackling projects where mathematics, performance, and reliability collide. From deep computational geometry to real-time imaging pipelines and AI breakthroughs, I bring algorithmic thinking into production-grade systems. What I Make Possible Computational Geometry & Simulation (CUDA): Architected a real-time electric field line simulator on GPUs, handling the dense math behind vector fields and parallel kernels. Delivered BIM-style geometry evaluation and Bezier-based drawing tools for clients who needed accuracy, interactivity, and stability at scale. Interactive Geometry Tools: Distinguished performer in delivering CAD-like solutions — room planning engines, floorplan canvases, and curve editors that match the sophistication of industry BIM systems. AI/ML (Computer Vision & OCR): Outperformed Tesseract and other baselines in image classification and OCR by building custom ML.NET and PyTorch pipelines. Leveraged trained models for chemistry-focused OCR, decoding molecular diagrams and classifying scientific data that generic engines couldn’t touch. MVVM & Real-Time Imaging: Built a live-view imaging pipeline in C#/WPF with sub-30ms latency, sustaining real-time overlays without freezing the UI by combining async/await with multithreading. I’m not just about ViewModels and data binding — I handle what most developers would call impossible to fold into MVVM, structuring complex rendering logic, continuous device streams, and performance-critical updates cleanly into responsive applications. Backend Architecture (ASP.NET Core): Built a search engine that transformed OCR’d exam content into queryable data for JEE/Olympiad problems, returning matches and solutions with precision. Secured APIs with Microsoft Identity + JWT for robust user management. XAML Engineering: Designed custom bindable controls — battery meters, temperature indicators, sound meters — powered directly by live camera and device data. Beyond templates, I create reusable controls that extend the XAML ecosystem. Algorithm Design: From extracting clean outlines from noisy contours to detecting bonds in chemical structures, I specialize in solving problems where raw data needs to become usable intelligence. Security & DevOps: Full-stack capability in authentication/authorization, user management APIs, and CI/CD pipelines to keep delivery seamless and production-ready. Beyond the Obvious If you can imagine extracting complex geometrical details from a live camera feed, training an AI model on that data, and rendering it in real-time with OpenGL/WebGL/WPF — I can deliver it or even make it interactable. This is just a mad example; the possibilities are much wider — not limited to camera or geometry, but really about what you can imagine when these skills combine, powered by my foundation in mathematics, computational geometry, CUDA, and AI. Recognition I am Expert-Vetted on Upwork, officially ranked among the top 920 freelancers worldwide in 2024 — a distinction far more selective than “top 1%.” This represents consistent success on high-stakes, algorithm-heavy projects with global clients. Skills & Tools C#, .NET, ASP.NET Core, REST APIs, Microsoft Identity, JWT Authentication, SignalR, EF Core, LINQ, MongoDB, Azure, Docker, Microservices WPF, WinUI 3, UWP, Uno Platform, MAUI, XAML, Custom Controls, MVVM, Real-Time Imaging, UI/UX for Desktop C++, CUDA, GPU Acceleration, High-Performance Computing (HPC), Parallel Computing, OpenGL, GLSL, CUDA-OpenGL Interoperability Computational Geometry, Physics Simulation, Algorithm Design, Contour Extraction, Data Structures & Optimization Machine Learning, Deep Learning, Computer Vision, OCR, Data-Driven Modeling, PyTorch, TorchVision Performance Optimization, Clean Architecture, Multithreading, Async/Await, CI/CD Pipelines, DevOps

Availability: Full-time freelancer, 𝟰𝟬+ hours/week, open to long-term collaborations. I’m a Full-Stack & AI Engineer with 10+ years of experience building web and mobile applications and 3+ years of specialized experience in AI and Large Language Models (LLMs). I design, develop, and deploy production-grade platforms, from scalable SaaS dashboards to AI-powered assistants, RAG systems, and voice agents. I work end-to-end: architecture → backend → frontend → cloud deployment, with a focus on clean code, maintainable systems, and high performance. Over the past few years, I’ve delivered solutions that integrate AI/LLM pipelines, vector search, real-time chat, and voice agents for enterprise and startup clients. 🤖 AI & LLM Expertise - MCP Server Development: Designing and integrating custom MCP servers for AI agents, enabling structured tool usage, external system integrations, database querying, and API orchestration. - Fine-Tuning: Persona creation, Q&A systems, and domain-specific models (medical, legal) using Mistral and Llama 3. - Synthetic Dataset Generation: Streamlining LLM training with high-quality datasets. - Evaluation Frameworks: Assessing LLM performance with custom metrics. - Cloud Deployment: Deploying LLMs on AWS and GCP. - AI Agents & Voice Bots: Proficient with LiveKit, Retail AI, OpenAI. - Open-Source Deployment: Expertise deploying models like vLLM on AWS/GCP/RunPod using SkyPilot. 🛠️ 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗺𝗲𝗻𝘁 𝗧𝗼𝗼𝗹𝘀 & 𝗙𝗿𝗮𝗺𝗲𝘄𝗼𝗿𝗸𝘀 ➜ LLM Tools: LangChain, Langsmith, Langfuse , Hugging Face, Transformers. ➜ Vector Databases: Chroma, FAISS, Pinecone, Qdrant , Opensearch ➜ AI Workflows: Flowise AI, LangFlow, StackAI. 🛠️ 𝗙𝘂𝗹𝗹 𝗦𝘁𝗮𝗰𝗸 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗺𝗲𝗻𝘁 𝗘𝘅𝗽𝗲𝗿𝘁𝗶𝘀𝗲 ➜ Languages & Frameworks: Python, Node.js, ReactJS. ➜ Database Management: MongoDB, MySQL, PostgreSQL , Supabase , FIrebase ➜ Frontend & Backend Integration: Seamlessly connecting APIs and user interfaces. 🌟 𝗔𝗱𝘃𝗮𝗻𝗰𝗲𝗱 𝗦𝗸𝗶𝗹𝗹𝘀 ➜ Open-Source LLMs: Proficiency in LLAMA 3, Mistral 7B, and Mixtral 8x7B. ➜ Prompt Engineering: Expertise in techniques like Chain of Thought, Few-shot Prompting, and Self-Reflection. ➜ Fast Inference: Implementing high-speed solutions with vLLM . 🌟 𝗪𝗵𝘆 𝗖𝗵𝗼𝗼𝘀𝗲 𝗠𝗲? With over 10 years of experience, I deliver scalable, cutting-edge solutions tailored to your project’s needs. Whether it's advanced AI models, MCP server development, LLM optimization, or full-stack development, I ensure top-notch results every time. Let’s collaborate to bring your ideas to life!

Custom-trained AI vision systems that detect manufacturing defects, count warehouse stock, protect retail floors from theft, and inspect garments for fabric flaws built for industrial buyers who need a vision model that actually works on their data, not a generic demo. I work with the Brainy Neurals team as the model-builder for industry-applied computer vision. My focus is the layer between "we have cameras" and "we have actionable detection" collecting the dataset, annotating it, training the model, evaluating it against real production conditions, shipping it as a working system. WHAT I BUILD Defect detection and quality inspection surface scratches, weld defects, packaging flaws, print errors, missing components, assembly mis-orientations, fabric defects, garment inspection. Custom-trained YOLO and segmentation models on your production images, evaluated on the edge cases that matter for your defect distribution. Retail analytics and loss prevention footfall counting, dwell-time heatmaps, queue analytics, planogram compliance, shelf-out detection, theft and shoplifting detection, restricted-zone intrusion, customer journey mapping. Runs on existing store cameras. Warehouse and logistics vision pallet counting, dock monitoring, SKU recognition, package counting on conveyors, forklift safety zones, PPE compliance (helmet, vest, gloves, harness), worker fall detection. Reduces shrinkage, audit exposure, safety incidents. Textile and apparel inspection fabric defect detection across woven, knit, and non-woven materials. Holes, stains, weave faults, print misregistration, colour variation, missing stitches, garment QC at finishing. Sports analytics player tracking with persistent IDs, ball detection, action recognition (shots, passes, fouls, rebounds), zone-aware scoring, jersey-digit recognition, highlight clips. Healthcare and patient safety fall detection in wards and assisted living, gait analysis, hand hygiene compliance, bed occupancy, restricted-area monitoring. Pose-based detection without face recognition where regulation requires. THE STACK I USE YOLOv11 and YOLOv12 when throughput matters; RT-DETR and Mask R-CNN when precision matters more than speed; SAM and Grounding DINO when datasets are small and zero-shot bootstrapping helps. Tracking is ByteTrack or DeepSORT depending on occlusion. Pose uses MediaPipe and custom keypoints. Foundation: PyTorch, OpenCV, ultralytics, timm, albumentations. Annotation and dataset work is where most CV projects quietly fail. I work with Roboflow, Label Studio, CVAT, and build custom tooling when off-the-shelf options corrupt formats. Datasets version-controlled. Edge cases sampled deliberately. Class imbalance handled with focal loss, augmentation, or class-aware sampling. Evaluation goes beyond mAP. I report precision, recall, F1, confusion matrices on rare classes, and false-positive rates at the operating threshold you care about. If your cost of a false positive differs from a false negative, the model is tuned for that asymmetry. WHO I BUILD FOR Manufacturing automotive, electronics, food and beverage, pharma packaging (blister packs, labels). Industry 4.0 quality, MES, OEE teams. Warehouses and 3PLs pallet count automation, dock-to-stock, shrinkage reduction, returns inspection. Retail fashion, grocery, big-box, convenience. Loss prevention, store ops, planogram, customer-experience teams. Textile and apparel fabric mills, garment factories, finishing units, brand QA. Sports broadcast, performance analytics, training instrumentation. Construction and AEC progress detection, structural element counting, site safety, drone inspection. Healthcare patient fall prevention, ward safety, hand hygiene, infection-control compliance. HOW I WORK Projects start with a 30-minute call where I look at sample images or video from your environment. I tell you whether a model can do what you want, dataset size needed, realistic accuracy ceiling, and where the model will fail. I would rather decline than over-promise on a dataset that cannot support it. Pricing is fixed-scope per milestone annotation, baseline training, evaluation, retraining, integration handoff. Hourly only for maintenance after shipping. THE BRAINY NEURALS BACKING I work with the Brainy Neurals team 15 AI engineers, NVIDIA Inception Partner, AWS Activate, Microsoft for Startups. When a project needs camera infrastructure, edge deployment, or multi-camera coordination, that capacity sits with the team I bring them in cleanly, you do not manage two vendors. For pure model-building I lead end to end myself. LET'S TALK IF You have images or video from your environment, you know what you want detected, and need a model trained on YOUR data not a generic detector that misses 30% of edge cases. Send a sample with your first message. I reply within 24 hours with a feasibility read, dataset size, timeline.

Recent work: - AI Automation - Document Review Product - My existing ComfyUI DiT video generation workflow (LTX-Video 2.3 I2V), which runs on H200. Optimized a native Blackwell (B200, sm_100) implementation of SageAttention specifically mirroring the SageAttention2++ architecture targetting CUDA 12.8+ - Product price analytics using data from Terapack, Pricecharting etc using NLP - Combined SAHI (Slicing Aided Hyper Inference) technique with Yolov5 to detect moving objects - A non-CNN-based motion vector estimator for image deblurring due to motion blur - Deep Learning based Document image processing solution - High Accuracy Face Tracking and face blurring solution Total Experience: +27 years Core Skills: - C/C++ - C# (Unity) - TensorFlow - Python - Flutter & Dart - NodeJS, Flask, Django - MongoDB - Git